Hi Robbie, I ran the tests again with a broker built from the latest trunk code and with a 0.16 client. The performance results are a lot better in almost every test setup (3-4 times better), except for Test #1 where I build up 100000 messages of unique grouping values and then see how many messages we can process in 5 minutes, using C++ mode (I got ~7800 messages before, and now ~5400). The overall comparisons are about the same, with C++ mode still performing significantly worse:
1. For the first test, for a regular queue and a default mode queue (with all messages having no grouping values, all having the same value, and all having different values), I am able to process 100000 messages in around 3 and a half minutes. In C++ mode, with all 100000 messages having different values, I can process only ~5400 in 5 minutes; with all having no value / the default group, ~44000 messages were processed in 5 minutes (after which I stop waiting and just clear the queue). 2. For the second test, for a default mode queue, the performance is again not affected by how many unprocessable messages we have at the front of the queue due to another consumer holding on to those messages' group for a long time. The time it takes to process 1000 messages of group B after N messages of group A (which is assigned to another consumer) is about the same as processing 1000 messages on a regular queue. With a C++ mode queue, the performance still gets worse the more messages of group A we have at the front of the queue. > Can I enquire what it is about your use case that would preclude use of a > higher prefetch? We want to avoid starving messages that can be processed by another consumer. We have multiple dequeue sessions, each listening to multiple queues. If we have a higher prefetch, then if there is a long-running message on a queue, the messages that were prefetched along with it will have to wait for a long time, even if another listener is available and can process those messages. And if we decide to add the message grouping configuration, we want to avoid starving a message of a particular group if there's an available listener -- if a listener can have 2 messages and 2 different groups assigned at a time, and processing of the first message takes a long time, then we might be unfairly starving the second message, even though other listeners can process it. We might be able to tolerate this to some extent if the prefetch is only 2, if this is the only way to improve overall throughput. I'll have to discuss this more with my team. Though this still wouldn't solve the problem with unique keys in C++ mode. > Also, you originally seemed to prefer the idea that > messages without a group header value would be not be grouped, so is there > a particular reason you are leaning towards using the shared groups > functionality which can't do that? > Basically, the ideal behavior we want is a third mode that combines the two. If no grouping value is specified, then treat it as though it has no group; for messages with a grouping value, ensure that only 1 listener can be processing messages from a particular group at a time, but don't tie a group to a particular consumer for the lifetime of that consumer. The reason we want the first part is that we'll have many messages that don't belong to any group, and we want them to be processed by different listeners in parallel. The reason we don't want a consumer to be associated with a group as long as the consumer lives is again to avoid starvation -- we don't want a consumer processing a message of group A for a long time result in starving group B's messages, because it happens so that the first group B message was processed by that consumer and so group B is assigned to the consumer, when there's another consumer of the same queue that is doing nothing and can process the B message. So we were thinking of asking you guys if you'd be open to introducing a third mode, mostly the same as the C++ mode but where no grouping value means no group, instead of the default group. Another workaround we thought of was for any message that doesn't belong to a group, we'll put its message ID in the grouping key property, so that essentially any listener can pick it up. That's one reason why we were testing the unique keys case, though it's also possible for us to get many different grouping values in a queue (though not at quite as high a number, e.g., 100000). But in C++ mode, with both unique values and same-group values at a high depth, we're seeing decreased performance anyway, so we might not be able to use it. Do you have any suggestions for what we should do? Are there any ideas you have for solving the performance issue with unique keys in C++ mode, so perhaps we could look into it more? Thanks, Helen > On 13 January 2014 20:34, Helen Kwong <helenkw...@gmail.com> wrote: > > > Hi Robbie, > > > > I am actually still running version 0.16 of the broker. It will take me a > > little time to set up the trunk code and rebuild and rerun the > experiments. > > Do you think your fix will likely make a difference? > > > > For the second case with a long-lived consumer being assigned the group > of > > many messages at the head of the queue, I was indeed using a prefetch of > 1. > > I ran it again (with version 0.16 still) with a prefetch of 2 as you > > suggested, and the dequeue time of the messages at the end was then not > > affected by the number of unprocessable messages at the beginning of the > > queue, about the same as the other test setups I ran. However, I think > > increasing prefetch to 2 might not work for our use case. > > > > For the first case with unique message groups, your explanation makes > sense > > and I think I understand it now. Do you think there is still a way to > > optimize this behavior, so that we don't need to possibly traverse > through > > the whole queue whenever a subscription is unassigned from a group? Since > > you mentioned that maintaining a fixed pointer to a queue entry would > > likely lead to memory retention issues, would having a weak reference be > a > > possible option? > > > > Thanks, > > Helen > > > > > > > > On Sat, Jan 11, 2014 at 7:58 AM, Robbie Gemmell < > robbie.gemm...@gmail.com > > >wrote: > > > > > Hi Helen, > > > > > > Can I check what version of the code you were using? I ask as the > latest > > > trunk or 0.26 release branch code is going to be necessary for > > correctness > > > and representative testing of the shared groups functionality, due to > the > > > defect fix I previously mentioned making recently. > > > You can find a nightly build of the trunk broker at: > > > > > > > > > https://builds.apache.org/view/M-R/view/Qpid/job/Qpid-Java-Artefact-Release/lastSuccessfulBuild/artifact/trunk/qpid/java/broker/release/and > > > I would need to build the 0.26 branch as the fix was introduced after > > > the latest RC. > > > > > > In your first case, I think the reason for the difference between the > > > default group and the unique group is also likely to be tied to the > > > 'findEarliestAssignedAvailableEntry' behaviour you mention later in > your > > > mail. For the default group case, that next message always going to be > a > > > message near the front of the queue. For the unique group, there isnt > > > actually going to be a message which matches, but it looks like it will > > > currently be checking every message to determine that and doing so > under > > > the synchronization, and thus probably preventing other deliveries > > > occurring at the time. That isnt a problem in the non-shared case > because > > > there isnt a need to synchronise the GroupManager as a whole, and even > > > going beyond that its also highly unlikely it would need to check as > many > > > messages before finding a match due to the signifcant difference in how > > > groups become associated with a particular subscription in the > non-shared > > > case. > > > > > > In your second case, your explanation seems likely and I think this > case > > > really reduces to just being a variant of the above behaviour. The > > > particular issue is that one could argue it shouldnt need to be doing > the > > > 'findEarliestAssignedAvailableEntry' task all that often in this case > if > > > you have a long-lived consumer, and so your mention of this makes me > > think > > > you are using a prefetch of 1. Using a prefetch of 1 currently means > that > > > the delivery state associated with the shared group effectively becomes > > > empty after each message, because messages are only fully added to the > > > group when they become acquired by a particular subscription, and they > > cant > > > be acquired until the previous message is consumed (or perhaps slightly > > > confusingly, explicitly not-consumed). If so, I expect it could be very > > > interesting to run this case again with a prefetch of 2 or more. The > > > obvious tradeoff with increasing prefetch is that a particular consumer > > > could then be assigned up to <prefetch> groups at a given point, though > > > likely not in your test case due to the large contiguous blocks of > > messages > > > for each group. > > > > > > I'm not sure that the suggestion to track the first message in the > group > > > would really work currently, due to the way the underlying queue data > > > structure works. Maintaining a fixed pointer into it like that is > likely > > to > > > lead to some undesirable memory retention issues, based on a related > but > > > far simpler case I fixed previously in a similar structure elsewhere in > > the > > > broker. Looking at the way messages become assigned to a group in the > > > shared group case may be a more viable path to handling your second > case > > > more gracefully. The unique groups from you first case would still need > > > something different though, as neither of these routes would really > help > > > there. > > > > > > Robbie > > > > > > On 11 January 2014 01:03, Helen Kwong <helenkw...@gmail.com> wrote: > > > > > > > Hi Robbie, > > > > > > > > I did some more testing to see whether message grouping will work for > > us, > > > > and compared the dequeue performance of a queue using message > grouping > > in > > > > default Java mode, a queue using C++ mode, and a queue not using > > message > > > > grouping. I found that when I use C++ mode, the performance can be > much > > > > worse than in other comparable setups, and was wondering if you could > > > help > > > > me understand why. > > > > > > > > 1. In one test, I have multiple listeners to a queue, enqueue 100000 > > > > messages to it, and see how many messages are processed in 5 > minutes. I > > > > have these different setups: > > > > > > > > - C++ mode queue with each message having a unique identifier > > > > > > > > - C++ mode queue with all messages having no grouping identifier (so > > all > > > > belong to the default group) > > > > > > > > - default mode queue with each message having a unique grouping > > > identifier > > > > > > > > - default mode queue with all messages having no grouping identifier > > > > > > > > - default mode queue with all messages having the same grouping > > > identifier > > > > > > > > - regular queue without a group header key configured > > > > > > > > All setups except for the first had about 35K - 39K messages > processed, > > > but > > > > for the first setup, there were under 8000 messages processed. What > > could > > > > explain this big difference? I’ve looked at the two grouping modes’ > > > > implementations of MessageGroupManager and see that C++ mode uses > > > > synchronized methods rather than a ConcurrentHashMap as in default > > mode, > > > so > > > > I’d guess there might be more contention because of that, but at the > > same > > > > time I can’t see why, in C++ mode, having a unique identifier for > each > > > > message results in throughput that is so much worse than having all > > > > messages in the default group. > > > > > > > > > > > > 2. I also wanted to see the impact of having many messages at the > head > > of > > > > the queue that a listener can’t process because the messages belong > to > > a > > > > group assigned to another consumer. E.g., have 10000 messages of > group > > A, > > > > followed by 1000 messages of group B, and listener 1 is holding on to > > the > > > > first A message for a long time -- see how long it will take > listener 2 > > > to > > > > process all the B messages. In this case C++ mode has performance > that > > > > degrades as the number of unprocessable group A messages at the front > > of > > > > the queue increases, whereas default mode's performance is > unaffected, > > > > about the same as processing 1000 messages on a regular queue. > > > > > > > > My rough guess from looking at DefinedGroupMessageGroupManager is > that > > > > whenever listener 2 is done with a group B message, the state change > > > > listener triggers the Group.subtract() to reset pointers for other > > > > subscriptions and consequently findEarliestAssignedAvailableEntry(). > > This > > > > then has to iterate through all the group A messages before it finds > > the > > > B > > > > message. Do you think this is the reason for the results I see? > > > > > > > > If so, is the idea here that other subscriptions of the queue could > > have > > > > skipped over the messages of a group while the group was assigned to > > some > > > > subscription S, so we need to tell them to set their pointers back? > If > > > that > > > > is indeed the idea, would it be possible to optimize it such that > when > > a > > > > group A is assigned to S and S gets its first message of the group, > we > > > > store what that first A message / queue entry is. Then when S is done > > > with > > > > the last A message, we can tell other subscriptions to go back to > that > > > > first entry, without having to iterate through the queue? > > > > > > > > Thanks a lot for your help! > > > > > > > > Helen > > > > > > > > > > > > On Tue, Jan 7, 2014 at 8:46 PM, Robbie Gemmell < > > robbie.gemm...@gmail.com > > > > >wrote: > > > > > > > > > ...and just to be super clear, though I think it it is mentioned > > > > correctly > > > > > in the docs this time, the 'default group' concept does not apply > in > > > the > > > > > regular / 'non shared' grouping mode. Messages that dont specify a > > > group > > > > > key value in that mode are simply not grouped in any way. > > > > > > > > > > On 8 January 2014 04:41, Robbie Gemmell <robbie.gemm...@gmail.com> > > > > wrote: > > > > > > > > > > > > > > > > > On 8 January 2014 04:33, Helen Kwong <helenkw...@gmail.com> > wrote: > > > > > > > > > > > >> Oh I see, I thought what you meant was that I could only alter > the > > > > > default > > > > > >> group in shared-groups mode starting with 0.24. > > > > > > > > > > > > > > > > > > No, I just missed that you said 0.16 and assumed 0.24 was the > > version > > > > you > > > > > > were using . You could always change it, just in more limited > ways > > in > > > > > > earlier releases. > > > > > > > > > > > > To make sure I'm > > > > > >> understanding this correctly -- changing the the default message > > > group > > > > > >> name > > > > > >> to something else in C++ mode won't change the serial processing > > > > > behavior > > > > > >> I > > > > > >> saw, right? > > > > > > > > > > > > > > > > > > Correct > > > > > > > > > > > > > > > > > >> Messages without a group identifier will still be considered to > > > > > >> be in a group -- rather than no group -- and they cannot be > > > processed > > > > by > > > > > >> multiple consumers concurrently? > > > > > >> > > > > > >> > > > > > > Yes. In the C++/shared-groups mode every message is considered to > > be > > > > in a > > > > > > group, it is just a case of whether the message specifies that > > group > > > or > > > > > > instead gets put into the default group. > > > > > > > > > > > > > > > > > > > > > > > >> Thanks, > > > > > >> Helen > > > > > >> > > > > > >> > > > > > >> On Tue, Jan 7, 2014 at 8:22 PM, Robbie Gemmell < > > > > > robbie.gemm...@gmail.com > > > > > >> >wrote: > > > > > >> > > > > > >> > I just noticed you said you were using 0.16, somehow glossed > > over > > > it > > > > > >> > originally and only noticed the 0.24 in the doc URL (its many > > > hours > > > > > past > > > > > >> > time I was asleep, I might be getting tired). > > > > > >> > > > > > > >> > Realising that, I should add that prior to 0.22 the only way > to > > > > alter > > > > > >> the > > > > > >> > default group in the shared-groups mode from 'qpid.no-group' > to > > > > > >> something > > > > > >> > else would have been via the 'qpid.default-message-group' > queue > > > > > declare > > > > > >> > argument when using an AMQP client to create the queue > > originally, > > > > and > > > > > >> for > > > > > >> > 0.22 itself only that and the system property approach I > > mentioned > > > > > would > > > > > >> > work. > > > > > >> > > > > > > >> > Robbie > > > > > >> > > > > > > >> > On 8 January 2014 04:03, Helen Kwong <helenkw...@gmail.com> > > > wrote: > > > > > >> > > > > > > >> > > Hi Robbie, > > > > > >> > > > > > > > >> > > I see. Thanks for the quick response and explanation! > > > > > >> > > > > > > > >> > > Helen > > > > > >> > > > > > > > >> > > > > > > > >> > > On Tue, Jan 7, 2014 at 7:43 PM, Robbie Gemmell < > > > > > >> robbie.gemm...@gmail.com > > > > > >> > > >wrote: > > > > > >> > > > > > > > >> > > > Hi Helen, > > > > > >> > > > > > > > > >> > > > The short answer to your question is that it is the > > > > documentation > > > > > >> which > > > > > >> > > is > > > > > >> > > > incorrect, and the behaviour you are seeing is expected. > > > > > >> > > > > > > > > >> > > > The long answer is, when that documentation was composed a > > > > segment > > > > > >> was > > > > > >> > > > missed out indicating this, and needs to be added to the > > docs. > > > > The > > > > > >> > > > behaviour listed for when no group is specified is only > true > > > of > > > > > the > > > > > >> > > > 'non-shared' groups supported by the Java broker, in the > > > > > C++/shared > > > > > >> > group > > > > > >> > > > mode any messages recieved without an explicit group value > > are > > > > all > > > > > >> > > assigned > > > > > >> > > > to a default group of 'qpid.no-group'. This is as per the > > > > > behaviour > > > > > >> of > > > > > >> > > the > > > > > >> > > > C++ broker itself, which is explained in the C++ broker > docs > > > at > > > > > the > > > > > >> end > > > > > >> > > of > > > > > >> > > > the following page > > > > > >> > > > > > > > > >> > > > > > > > > >> > > > > > > > >> > > > > > > >> > > > > > > > > > > > > > > > http://qpid.apache.org/releases/qpid-0.24/cpp-broker/book/Using-message-groups.html > > > > > >> > > > . > > > > > >> > > > For the 0.24 Java broker, this default shared group can be > > > > changed > > > > > >> > > > broker-wide using the Java system property > > > > > >> > > > 'qpid.broker_default-shared-message-group', or can be > > > overriden > > > > > for > > > > > >> an > > > > > >> > > > individual queue during creation programatically via AMQP > > > > clients > > > > > or > > > > > >> > the > > > > > >> > > > management interfaces through use of the argument > > > > > >> > > > 'qpid.default-message-group' or > 'messageGroupDefaultGroup'. > > > > > >> > > > > > > > > >> > > > I coincidentally happened to have fixed a defect with the > > > shared > > > > > >> groups > > > > > >> > > > functionality last night on trunk. Its not yet included in > > the > > > > > >> imminent > > > > > >> > > > 0.26 release, though I am about to request whether that is > > > > > possible. > > > > > >> > > > https://issues.apache.org/jira/browse/QPID-5450 > > > > > >> > > > > > > > > >> > > > Robbie > > > > > >> > > > > > > > > >> > > > On 8 January 2014 02:43, Helen Kwong < > helenkw...@gmail.com> > > > > > wrote: > > > > > >> > > > > > > > > >> > > > > Hi, > > > > > >> > > > > > > > > > >> > > > > I use the Java broker and client, version 0.16, and am > > > > > considering > > > > > >> > > using > > > > > >> > > > > the message grouping feature ( > > > > > >> > > > > > > > > > >> > > > > > > > > > >> > > > > > > > > >> > > > > > > > >> > > > > > > >> > > > > > > > > > > > > > > > http://qpid.apache.org/releases/qpid-0.24/java-broker/book/Java-Broker-Queues.html#Java-Broker-Queues-OtherTypes-Message-Grouping > > > > > >> > > > > ). > > > > > >> > > > > From testing I've done, there seems to be a bug with the > > C++ > > > > > >> > > > compatibility > > > > > >> > > > > model, and I'm wondering if this is a known issue. > > > > Specifically, > > > > > >> in > > > > > >> > my > > > > > >> > > > test > > > > > >> > > > > I have a queue configured to use a group header field > with > > > > > >> > > > > "qpid.group_header_key" and C++ mode with > > > > > "qpid.shared_msg_group", > > > > > >> > and > > > > > >> > > > have > > > > > >> > > > > multiple listeners to the queue. Each listener will > sleep > > > for > > > > a > > > > > >> short > > > > > >> > > > > amount of time when it receives a message before > > returning. > > > I > > > > > then > > > > > >> > > > enqueue > > > > > >> > > > > 10 messages that do not have a value in the group header > > > field > > > > > to > > > > > >> the > > > > > >> > > > > queue. Since the doc says that messages without a value > in > > > the > > > > > >> > grouping > > > > > >> > > > > header will be delivered to any available consumer, the > > > > > behavior I > > > > > >> > > expect > > > > > >> > > > > is that the messages will be processed in parallel, > i.e., > > > when > > > > > >> > > listener 1 > > > > > >> > > > > is holding on to a message and sleeping, listener 2 can > > > > receive > > > > > >> > another > > > > > >> > > > > message from the queue. But what I see is that the > > messages > > > > are > > > > > >> > > processed > > > > > >> > > > > serially -- message 2 won't be received by some thread > > until > > > > > >> message > > > > > >> > 1 > > > > > >> > > is > > > > > >> > > > > done. When I use the default mode instead of C++ mode, > > then > > > I > > > > > get > > > > > >> the > > > > > >> > > > > parallel processing behavior. > > > > > >> > > > > > > > > > >> > > > > Is this is a known bug, and is there a fix for it > already? > > > > > >> > > > > > > > > > >> > > > > Thanks, > > > > > >> > > > > Helen > > > > > >> > > > > > > > > > >> > > > > > > > > >> > > > > > > > >> > > > > > > >> > > > > > > > > > > > > > > > > > > > > > > > > > > >
